Google DeepMind's EmbeddingGemma 2 embeds four modalities with 740M Apache-licensed parameters

The 740M figure is a ceiling. The text backbone is 270M parameters, the vision encoder 170M and the audio encoder 300M ^[5]. The two media encoders are independent and load only when needed ^[6]. A text-only index therefore loads about 36% of the model ^[20], and a text-plus-image index loads 440M ^[18]. Audio is the expensive modality. Its encoder alone is about 41% of the total, larger than the whole text stack ^[19].

How an encoder gets dropped depends on the runtime. In sentence-transformers it goes through `config_kwargs`, and the card says the method differs among model libraries ^[9]. The smaller footprint exists only if the library a team ships can skip the audio weights.

Storage gets the same modular treatment. Matryoshka training lets the 768-dimension vector be cut to 512, 256 or 128 by keeping its leading dimensions and re-normalising ^[2]. The card advertises up to a 6x reduction in vector storage ^[1]. That figure is 768 divided by 128 ^[16]. The same card says quality loss is minimal only down to 256 dimensions, and that 128 is best suited to text-only workloads ^[2]. For an index holding images, video or audio, the saving the card supports is 3x ^[17]. We think the truncation design is the most useful engineering on the card for anyone running an index at scale. Because the cut keeps leading dimensions, vectors stored at 768 can be shortened later without re-embedding the corpus ^[2].

Text inputs expect a short task prefix. The model was trained with them, and leaving them off still works but costs precision; images, video and audio go in without one ^[8]. Retrieval uses separate query and document prefixes, while classification and similarity apply one prefix to every input ^[10]. One default deserves attention. `prompt_name="Document"` applies `title: none`, so a corpus with real titles has to be formatted by hand as `title: {title} | text: {content}` ^[11]. A pipeline that relies on the named prompt drops every title it was given, and nothing in the output says so ^[11].

On hosted APIs, the evidence is thin. The card does not include prices or scores against hosted embedding services. Every reported result comes from the full-precision checkpoint ^[12], while the model is pitched at phones and laptops ^[14]. For those scores to carry over, the build a team actually ships has to land close to full precision. The team's corpus also has to resemble the text, code, vision, visual-document, video and audio benchmarks Google DeepMind used ^[13]. Code is the one area where the card gives a delta, a roughly 14% gain over the predecessor, alongside support for more than 100 languages ^[7]. In our view the Apache 2.0 licence ^[3] and the loadable-encoder design make a bake-off against a hosted API cheap enough to run before the next contract renewal.